YouTube transcripts → Claude_
Claude 3.5 Sonnet and Opus have a 200k-token context window — big enough for almost any single YouTube video. Here's how to get a transcript in, structure the prompt so Claude actually follows it, and scale the workflow across a full back-catalog.
1. Pull the transcript
Open the transcript extractor, paste the YouTube URL, and export the transcript. For Claude, plain text without timestamps is best — it saves tokens and doesn't distract the model.
2. Wrap it in XML tags
Claude's system prompt is trained to follow XML-style delimiters more reliably than Markdown or plain text. Wrap the transcript in <transcript>...</transcript> and reference the tag by name in your instructions.
3. Prompt templates
You are a research assistant. Read the transcript inside <transcript> tags and return:
1. A 3-sentence TL;DR
2. Five key takeaways as bullets
3. Any numbers, dates, or named entities as a table
<transcript>
{paste transcript here}
</transcript>Extract every direct quote from the speaker that would work as a pull-quote. Return JSON: [{ "quote": "...", "context": "..." }]. Do not paraphrase.
<transcript>
{paste transcript here}
</transcript>List every factual claim in the transcript. For each, mark: verifiable | opinion | unclear. Return a Markdown table with columns Claim | Type | Timestamp.
<transcript>
{paste transcript here}
</transcript>4. Batch across your library
For Claude workflows at scale, use the batch pipeline to send up to 1,000 transcripts through the same prompt in a single job — ideal for building a searchable Claude-powered knowledge base from an entire podcast or lecture series.
Also see the ChatGPT & Claude repurposing guide for chunking strategy when you want to run the same prompt across multiple models.